Elastomagnetic sensing offers a non-destructive route to estimate axial tension in steel rods, but amplitude-only practices are easily biased by temperature drift, baseline offsets, and measurement noise. We address this by extracting a log-amplitude ratio from secondary-coil voltages referenced to each sensor’s unloaded reference state, and by explicitly modeling thermal influences with simple thermal indicators derived from ambient temperature and secondary-coil resistance. Raw waveforms are band-pass filtered around the excitation and converted to complex phasors. The log-amplitude ratio is robustly regressed against temperature change and normalized resistance change to quantify thermal sensitivity; the resulting coefficients are then used to subtract thermal effects from the amplitude feature. The corrected feature is standardized, projected to a single latent coordinate via robust weighted least squares, and then mapped to tension using a monotone piecewise cubic Hermite interpolating polynomial (PCHIP), which preserves the expected monotonic relation while avoiding overfitting. Time series of estimated tension for each sensor are produced directly. To keep thermal indicators comparable across datasets and devices, we apply an internal linear normalization of the resistance–temperature relation without altering the underlying measurements. Laboratory tension/temperature campaigns and field deployment show that amplitude with thermal decoupling markedly improves stability and accuracy over amplitude-only baselines, enabling practical continuous monitoring of prestressed steel rods.